Training of Joint Depth Prediction and Completion

2. The depth system of claim 1, wherein the network module includes instructions to generate the depth loss and the image loss including instructions to separately compute a sum of variances and a weighted squared mean of errors in log space and combine the depth loss and the image loss into a combined loss to update the depth model.

3. The depth system of claim 1, wherein the network module includes instructions to update the depth model including instructions to train the depth model by using the depth loss and the image loss to adapt weights of the SAN, an image encoder, and a depth decoder of the depth model.

4. The depth system of claim 1, wherein the network module includes instructions to generate the first depth map and the second depth map including instructions to execute the depth model twice by activating the SAN when generating the second depth map and deactivating the SAN when generating the first depth map.

5. The depth system of claim 4, wherein the network module includes instructions to activate the SAN when generating the second depth map including instructions to process the sparse depth data with the SAN to generate the depth features and inject the depth features into a depth decoder of the depth model to facilitate generating the second depth map.

6. The depth system of claim 1, the network module includes instructions to acquire training data that includes the monocular image and the sparse depth data from at least one sensor of a device.

7. The depth system of claim 1, wherein the network module includes instructions to provide the depth model to generate depth maps to control a device to navigate through a surrounding environment according to the depth map that identifies distances to objects in the surrounding environment.

8. The depth system of claim 1, wherein the network module includes instructions to update the depth model including instructions to determine learned weights for the depth features and image features applied when concatenating the depth features and the image features via skip connections of the depth model.

10. The non-transitory computer-readable medium of claim 9, wherein the instructions to generate the depth loss and the image loss include instructions to separately compute a sum of variances and a weighted squared mean of errors in log space and combine the depth loss and the image loss into a combined loss to update the depth model.

11. The non-transitory computer-readable medium of claim 9, wherein the instructions to update the depth model include instructions to train the depth model by using the depth loss and the image loss to adapt weights of the SAN, an image encoder, and a depth decoder of the depth model.

12. The non-transitory computer-readable medium of claim 9, wherein the instructions to generate the first depth map and the second depth map include instructions to execute the depth model twice by activating the SAN when generating the second depth map and deactivating the SAN when generating the first depth map.

13. The non-transitory computer-readable medium of claim 9, wherein the instructions to activate the SAN when generating the second depth map include instructions to process the sparse depth data with the SAN to generate the depth features and inject the depth features into a depth decoder of the depth model to facilitate generating the second depth map.

15. The method of claim 14, wherein generating the depth loss and the image loss includes separately computing a sum of variances and a weighted squared mean of errors in log space and combining the depth loss and the image loss into a combined loss to update the depth model.

16. The method of claim 14, wherein updating the depth model includes training the depth model by using the depth loss and the image loss to adapt weights of the SAN, an image encoder, and a depth decoder of the depth model.

17. The method of claim 14, wherein generating the first depth map and the second depth map includes executing the depth model twice by activating the SAN when generating the second depth map and deactivating the SAN when generating the first depth map.

18. The method of claim 17, wherein activating the SAN when generating the second depth map includes processing the sparse depth data with the SAN to generate the depth features and injecting the depth features into a depth decoder of the depth model to facilitate generating the second depth map.